Calculation of fixed effects dummy variable coefficients

[KeithWoodward]

Hi Gareth,

I'm trying to figure out how eviews does fixed effects (cross section and period effects). I'm using eviews 7, build 'Dec 8 2009'.

Specifically, how does eviews calculate the dummy variable co-efficients which are displayed when I select View -> Fixed/Random effects -> Period effects, after estimating a panel data regression with cross-section and period effects? Unusually, there are (n) dummies, rather than (n-1), even though I include a constant (c) in the regression.

If I manually include fixed effects by estimating a new regression with a dummy variable for each entity and period, then I get the same co-efficients and standard errors on the original variables, but the constant (c) is different and the co-efficients on the fixed effect dummies are all different to the automatically generated fixed effects displayed by eviews when the steps in the above paragraph are followed and the results are compared.

It's clear that eviews adjusts the regression's constant and fixed effects in some way, and I'd like to know how. These are the only clues that I was able to find:

http://forums.eviews.com/viewtopic.php?f=4&t=4417
"However, what EViews does is express each country's fixed effect as the difference from the over all intercept, rather than as an individual intercept. Doing it this way means there is no multicollinearity problem."

http://www.fordham.edu/economics/mcleod ... edData.pdf
"Note that when you select a fixed or random effects specification, EViews will automatically add a constant to the common coefficients portion of the specification if necessary, to ensure that the effects sum to zero."

Thanks a lot for your help. I'm teaching a few classes of students how to do regressions on eviews and these problems have been a stumbling block so any help would be greatly appreciated.
Best regards,
Keith Woodward
Teacher, UTS Sydney, Australia

[EViews Gareth]

It is just as your first quote describes. Rather than giving the full intercept for each cross-section, we give the deviation from the overall intercept (or, in other words, the deviation from the mean of the fixed effects, where the mean is given by the constant coefficient).

[KeithWoodward]

Thanks for your reply Gareth, much appreciated. I have two clarifying questions.

1. Is the 'overall intercept' the constant in the regression?

2. Should the entity (cross section) fixed effects sum to zero, and the time (period) fixed effects also sum to zero?

Thank you,
Keith

[EViews Glenn]

Yes.

In general, not necessarily, though in the balanced design case, yes.

[KeithWoodward]

Thanks Glenn, that helps a lot.

I am trying to reconcile the difference between the eviews output and the dummy-variable approach that we teach students.

I was going to ask how eviews does fixed effects under-the-hood, but thought of looking at some other projects first to see how they do fixed effects. I found this:
http://www.stata.com/support/faqs/stat/xtreg2.html
Does eviews work similarly?

Thanks,
Keith

[EViews Glenn]

My quick look says that's right...

[KeithWoodward]

Great, thanks Glenn.

[Arkate]

Hello.

I would like to ask for a clarification since it's the first time I'm trying to use the fixed effects model.
Specifically, from what I have understood in order to obtain the coefficients of the model without an intercept, I should subtract the coefficients that eviews gives from the intercept. Is that correct? If not, how could I obtain a cross section fixed effects model without an intercept?

Thank you.

[EViews Gareth]

Correct

[kbandesz]

Hi,

I estimated a fixed effect regression on an unbalanced panel, and I noticed that the sum of the effects is not zero (very far from zero). Earlier in this thread, Glenn hinted that the sum of the FEs is "In general, not necessarily [zero] , though in the balanced design case, yes." So I reckon the issue in my case is the unbalanced nature of the panel. Am I right? If so, could you give me some pointer why this is the case? What is the restriction on the FEs to identify them separately from the constant?

Thanks,
Andras